demerzel-compound

name	demerzel-compound
description	Run a meta-compounding cycle — scan evolution log, detect promotions/demotions, assess governance effectiveness, propose self-improvements

Demerzel Meta-Compounding Cycle

Run a full governance compounding cycle. This is how Demerzel improves her own governance — recursively, based on evidence.

Usage

/demerzel compound — run full cycle /demerzel compound scan — just scan for candidates without proposing changes /demerzel compound promote — show promotion candidates only /demerzel compound waste — show deprecation/waste candidates only

The Compounding Cycle

0. Liveness check — is the evolution log fresh?
       ↓ (if stale, harvest first)
1. Scan evolution log
       ↓
2. Detect promotion candidates (pattern → policy → constitutional)
       ↓
3. Detect deprecation candidates (unused, stale, redundant)
       ↓
4. Assess confidence calibration accuracy
       ↓
5. Evaluate governance effectiveness per artifact
       ↓
6. Propose self-improvements (Kaizen proactive/innovative)
       ↓
7. Log results back to evolution log (the loop compounds itself)

Step 0: Liveness Check (MANDATORY)

Before reading the evolution log, verify it is fresh enough to compound from. Discussion #242 (2026-04-30) demonstrated the failure mode: the cycle ran on 27-day-old state and produced static inventory while 130+ commits across Demerzel + ix + hari were invisible to it.

Procedure:

1. Find the newest mtime across state/evolution/*.json — call it T_log.
2. Run git log --since="$T_log" --oneline in the current repo. Repeat for sibling repos at ~/source/repos/{ix,hari,tars,ga} (skip any that are not present).
3. Decision:
   • If no new commits in any repo → log is fresh, proceed to Step 1.
   • If new commits exist but T_log is < 7 days old → log is current enough, proceed but flag the activity for inclusion.
   • If T_log ≥ 7 days old AND new commits exist → STALE-STATE WARNING. Do not produce a normal compound report. Either:
     • (a) Trigger /demerzel harvest first, then re-enter Step 0, OR
     • (b) If harvest is unavailable, fall back to a manual compound: read the git log + recent session memories from ~/.claude/projects/*/memory/ + sibling-repo git logs, and write the deltas as new evolution entries. The 2026-04-30 corrective cycle (docs/reports/2026-04-30-compound-corrective.md) is the reference template for this fallback.

The cycle's own correctness depends on its inputs. Treat input freshness as a first-class precondition, not an assumption.

Step 1: Scan Evolution Log

Read all files from state/evolution/ — this is Demerzel's live evolution log. Fall back to examples/sample-data/governance-evolution-seed.json only if state/evolution/ is empty.

For each artifact, check:

• citation_count — how often cited in governance decisions
• violation_count — how often violated
• compliance_rate — violations / total assessments
• last_cited — when was it last relevant?
• events — what has happened to this artifact?

Step 2: Detect Promotion Candidates

Pattern → Policy promotion (evidence-based):

• Look for operational patterns appearing in 3+ PDCA cycles, reconnaissance findings, or compliance decisions
• Check: is there a governance gap that a new policy would fill?
• Threshold: 3+ appearances with measurable positive impact

Policy → Constitutional promotion (human judgment required):

• Look for policies with: 100% compliance rate, high citation count, sustained effectiveness
• Check: would violation of this policy constitute fundamental harm?
• Threshold: sustained inviolability + no exceptions in evolution log
• NOTE: This only generates a proposal — human must approve

Step 3: Detect Deprecation Candidates

Apply Kaizen waste taxonomy:

• Ceremony without value: Artifacts with zero citations for 90+ days
• Redundant governance: Multiple artifacts covering the same concern
• Stale artifacts: Artifacts referencing deprecated capabilities
• Over-engineering: Policies with unused provisions

Flag but do NOT auto-deprecate. Generate a proposal for review.

Step 4: Assess Confidence Calibration

Review recent governance decisions for calibration accuracy:

• Were high-confidence decisions actually correct?
• Were low-confidence decisions appropriately escalated?
• Is any agent systematically over- or under-confident?

Uses the confidence calibration protocol from policies/scientific-objectivity-policy.yaml.

Step 5: Evaluate Governance Effectiveness

For each artifact, update the assessment:

• effectiveness: T (clearly effective), F (ineffective), U (not enough data), C (mixed results)
• recommendation: maintain, promote, demote, deprecate, investigate

Use tetravalent logic — don't force binary judgments:

• Unknown (U) is fine for new artifacts — it means "gather more data"
• Contradictory (C) is important to surface — don't average away mixed results

Step 6: Propose Self-Improvements

Based on findings, propose concrete improvements:

Proactive Kaizen (Demerzel can propose, skeptical-auditor reviews):

• New policies for frequently-used patterns
• Policy refinements based on usage data
• Schema updates for better validation

Innovative Kaizen (requires human authorization):

• Constitutional amendments for proven-inviolable policies
• Structural changes to the governance framework
• New personas for ungoverned capabilities

Each proposal follows PDCA:

• Plan: hypothesis about what to improve
• Do: smallest testable change
• Check: measure impact via evolution log
• Act: standardize or revert

Step 7: Log Results

Update the evolution log with:

• New events (type: "compounding_cycle") on each assessed artifact
• Updated metrics (citation counts, compliance rates)
• Updated assessments (effectiveness, recommendations)
• Proposals generated (with evidence and confidence)

This is the recursive part — the compounding cycle's own results feed into the next cycle.

Output Format

=== Demerzel Meta-Compounding Cycle ===
Date: [today]

PROMOTION CANDIDATES:
  [list with evidence]

DEPRECATION CANDIDATES:
  [list with evidence]

CALIBRATION ASSESSMENT:
  [accuracy report]

EFFECTIVENESS SUMMARY:
  | Artifact | Effectiveness | Recommendation |
  [table]

PROPOSED IMPROVEMENTS:
  [list with PDCA format]

CYCLE LOGGED: [count] artifacts updated

Scheduling

Run after the weekly harvest — the harvest provides fresh data for the compounding cycle.

Add to Streeling workflow or run manually:

CronCreate("47 13 * * 5", "/demerzel compound")

(Runs Fridays at 8:47 AM EST, 30 minutes after the harvest)

Integration

• Reads from: governance evolution log, Kaizen waste taxonomy, confidence calibration protocol
• Writes to: governance evolution log (updated assessments and events)
• Feeds into: /demerzel promote for promotion proposals, /demerzel evolve for evolution insights
• Governed by: Asimov constitution (can't self-modify constitutions), Demerzel mandate (self-governance requires external review)

Constitutional Constraint

Demerzel CANNOT approve her own constitutional changes. Per the mandate:

• Self-governance requires skeptical-auditor review (routine)
• Constitutional promotions require human approval (always)
• The compounding cycle proposes — it never unilaterally changes governance

State Maintenance (MANDATORY)

The compounding cycle is the primary writer of evolution state. It MUST:

Before Cycling

1. Read ALL files from state/evolution/ — this is the input
2. Read recent beliefs from state/beliefs/ for context
3. Read any PDCA records from state/pdca/ for improvement tracking

After Cycling

1. Update evolution files in state/evolution/ — add new events, update metrics, refresh assessments
2. Create PDCA records in state/pdca/ for each proposed improvement
3. Update beliefs in state/beliefs/ if the cycle revealed new truths
4. Never delete evolution entries — append events, update metrics in place

The Compound Loop

Each cycle's output is the next cycle's input. The evolution log grows richer over time:

• Cycle N finds alignment-policy has 3 citations → marks as promotion candidate
• Cycle N+1 sees the promotion candidate flag → evaluates if threshold sustained
• Cycle N+2 either promotes or downgrades based on accumulated evidence

Source

logic/governance-evolution.schema.json, policies/scientific-objectivity-policy.yaml (confidence calibration), policies/kaizen-policy.yaml (waste taxonomy, PDCA), constitutions/demerzel-mandate.md Section 4 (accountability)